hud-python 0.6.2__tar.gz → 0.6.4__tar.gz

{hud_python-0.6.2 → hud_python-0.6.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hud-python
-Version: 0.6.2
+Version: 0.6.4
 Summary: SDK for the HUD platform.
 Project-URL: Homepage, https://github.com/hud-evals/hud-python
 Project-URL: Bug Tracker, https://github.com/hud-evals/hud-python/issues

hud_python-0.6.4/cookbooks/connect4-selfplay/README.md

@@ -0,0 +1,57 @@
+# Connect Four self-play
+
+Symmetric self-play RL on a 6×7 Connect Four board. Draws are rare (you need a
+full 42-cell board with no four-in-a-row), so the win/loss reward signal
+persists as the policy improves and the GRPO advantage stays non-zero.
+
+## How it works
+
+- One agent ("outer") plays a full game against an inner model on the **same
+  slug** — true self-play. `seed % 2` decides who drops first, for symmetric
+  first-move coverage.
+- Each game trains **both sides at once**: the outer agent's `Run` (reward from
+  its perspective) plus a hand-built `TrajectoryPayload` for the inner model
+  with the flipped reward (`1 - outer_reward`).
+- `group_size=2` pairs each game's two trajectories so the GRPO advantage is
+  `reward - 0.5` per game.
+- `loss_fn="ppo"` clips the importance-sampling ratio, so a single lucky game
+  can't blow up the update.
+
+The training loop uses the public API directly — `forward_backward` accepts
+`Run` and `TrajectoryPayload` mixed, so no private helpers are needed.
+
+## Setup
+
+```bash
+hud models fork Qwen/Qwen3.5-4B --name c4-selfplay   # prints a slug like c4-selfplay-<id>
+```
+
+Put your `HUD_API_KEY` in a `.env` here (or the environment).
+
+## Run
+
+Local sanity check (one game, cheap external model as the outer agent):
+
+```bash
+hud eval env.py claude --model claude-haiku-4-5
+```
+
+Train:
+
+```bash
+python train.py --model c4-selfplay-<id> --steps 20 --group 4 --lr 1e-5
+```
+
+## Tuning notes
+
+- **Memory scales with `tasks × group`.** Each task×rollout is a fresh `env.py`
+  subprocess. With 8 tasks and `--group 4` that's 32 concurrent games. Connect
+  Four games can run up to 42 plies, so they cost more tokens and time per game —
+  start at `--group 4` and raise only if you have RAM headroom.
+- **Watch the server-side metrics.** The loop prints local win/draw/loss counts
+  each step and the last few checkpoints' `mean_reward` / `reward_std` via
+  `trainer.checkpoints()` at the end. A healthy run keeps non-trivial
+  `reward_std` (within-group spread); if it collapses, the policy has saturated.
+- **Reset on changes.** If you edit the reward or the board, roll the head back
+  to a clean checkpoint (`hud models head <slug> --set <id>`) or fork fresh —
+  don't keep training a policy shaped by the old objective.

hud_python-0.6.4/cookbooks/fireworks-rl-training/README.md

@@ -0,0 +1,114 @@
+# Fireworks RL Training
+
+Direct Fireworks Training API loop over the same arithmetic preview task used by
+`cookbooks/rl-training`.
+
+This does **not** use Fireworks native datasets or RFT jobs. It follows the
+Training API service path from the Fireworks docs:
+
+1. `FiretitanServiceClient.from_firetitan_config(...)`
+2. `create_deployment_sampler(...)` for high-parallel rollouts
+3. local grading of HUD-style multiplication tasks
+4. `forward_backward_custom(...)` + `optim_step(...)`
+5. `save_weights_for_sampler(...)` + sampler refresh
+
+References:
+
+- Fireworks Training API introduction: https://docs.fireworks.ai/fine-tuning/training-api/introduction
+- Training and sampling lifecycle: https://docs.fireworks.ai/fine-tuning/training-api/training-and-sampling
+- Loss functions / GRPO reference: https://docs.fireworks.ai/fine-tuning/training-api/loss-functions
+
+## Setup
+
+The repo-level `.env` is loaded automatically. It must contain:
+
+```bash
+FIREWORKS_API_KEY=...
+FIREWORKS_ACCOUNT_ID=...
+```
+
+Install the isolated cookbook environment:
+
+```bash
+uv sync --pre
+```
+
+## Calibrate task difficulty first
+
+Calibration defaults to Fireworks' OpenAI-compatible inference API, so it does
+**not** create a trainer, provision a Training API deployment, or call
+`optim_step`. This is the cheap way to tune task difficulty before paying for a
+Training API run.
+
+The calibration model is separate from the training base model because the
+`lorenss` key currently exposes only a small serverless inference catalog (no
+Qwen3 8B deployment). Override it with `--inference-model` if you have a closer
+deployed model.
+
+```bash
+uv run train.py --calibrate-only --groups-per-step 8 --rollouts-per-prompt 8 --parallelism 32
+```
+
+The goal is a reward distribution with variance. If reward is all zero, make the
+task easier:
+
+```bash
+uv run train.py --calibrate-only --min-a 10 --max-a 99 --min-b 2 --max-b 9
+```
+
+If reward is all one, make the task harder:
+
+```bash
+uv run train.py --calibrate-only --min-a 1000 --max-a 9999 --min-b 11 --max-b 99
+```
+
+The current defaults are calibrated for the visible `gpt-oss-120b` inference
+model on the `lorenss` key: 2-digit by 1-digit multiplication with a direct
+"reply only with the integer" prompt. A 32-rollout calibration gave a non-trivial
+baseline (`reward_mean ~= 0.22`, `reward_std ~= 0.42`), while the original
+3-digit by 2-digit range was all-zero.
+
+## Train
+
+Once calibration has non-trivial rewards:
+
+```bash
+uv run train.py --steps 5 --groups-per-step 8 --rollouts-per-prompt 8 --parallelism 32
+```
+
+This uses the direct Training API managed service path. If you want calibration
+to go through the managed deployment sampler too, pass
+`--calibration-backend managed`; this provisions the same resources as training.
+
+### Current Fireworks preview account blocker
+
+On the `lorenss` preview account, trainer creation currently fails before the
+first train step with:
+
+```text
+failed to ensure FIREWORKS_API_KEY secret: unkey inference api id is not configured
+```
+
+This happens even with `create_deployment=False`, so it is an account/control
+plane provisioning issue rather than a problem in the rollout or loss code. Once
+Fireworks enables the missing Unkey inference API config for the account, the
+same `uv run train.py ...` command should proceed to trainer startup and the
+first `forward_backward_custom(...)` call.
+
+Metrics are written to:
+
+- `runs/fireworks-rl-preview/metrics.jsonl`
+- `runs/fireworks-rl-preview/reward_loss.png` if `matplotlib` is installed
+
+## Notes
+
+- Defaults use Qwen 3 8B full-parameter training:
+  - `accounts/fireworks/models/qwen3-8b`
+  - `Qwen/Qwen3-8B`
+  - `accounts/fireworks/trainingShapes/qwen3-8b-128k`
+- LoRA can be tested with `--lora-rank N`, but the validated Qwen3 8B training
+  shape currently rejects LoRA mode on the `lorenss` preview account.
+- The first checkpoint sync happens after step 0 and subsequent rollouts sample
+  the updated weights through the same deployment.
+- `--keep-trainer` and `--keep-deployment` are available for debugging. By
+  default the trainer is cleaned up and the deployment scales to zero on exit.

hud_python-0.6.4/cookbooks/fireworks-rl-training/pyproject.toml

@@ -0,0 +1,19 @@
+[project]
+name = "fireworks-rl-training"
+version = "0.1.0"
+description = "Direct Fireworks Training API RL loop over HUD-style arithmetic tasks"
+requires-python = ">=3.11,<3.13"
+dependencies = [
+    "fireworks-ai[training]",
+    "hud-python",
+    "matplotlib",
+    "python-dotenv",
+    "torch>=2",
+    "transformers>=4.55",
+]
+
+[tool.uv]
+package = false
+
+[tool.uv.sources]
+hud-python = { path = "../..", editable = true }

{hud_python-0.6.2 → hud_python-0.6.4}/hud/agents/__init__.py

@@ -8,7 +8,12 @@ from __future__ import annotations
 from typing import TYPE_CHECKING, Any, cast
 
 from hud.types import AgentType
-from hud.utils.gateway import build_gateway_client, list_gateway_models
+from hud.utils.gateway import (
+    build_gateway_client,
+    gateway_model_aliases,
+    list_gateway_models,
+    normalize_gateway_model_id,
+)
 
 if TYPE_CHECKING:
     from typing import TypeAlias
@@ -27,6 +32,8 @@ def create_agent(model: str, **kwargs: Any) -> GatewayAgent:
 
     For direct API access with provider API keys, instantiate the agent classes directly.
     """
+    requested_model = model
+    model = normalize_gateway_model_id(model)
     agent_type = next((candidate for candidate in AgentType if candidate.value == model), None)
     if agent_type is not None:
         model_id = model
@@ -73,7 +80,8 @@ def create_agent(model: str, **kwargs: Any) -> GatewayAgent:
                 for n in (gm.id, gm.name, gm.model_name)
                 if isinstance(n, str)
             ]
-            near = difflib.get_close_matches(model, known, n=3, cutoff=0.5)
+            known.extend(gateway_model_aliases())
+            near = difflib.get_close_matches(requested_model, known, n=3, cutoff=0.5)
             hint = (
                 f" Did you mean: {', '.join(near)}?"
                 if near
@@ -84,7 +92,7 @@ def create_agent(model: str, **kwargs: Any) -> GatewayAgent:
                 if gateway_models
                 else "the HUD gateway registry (empty — is HUD_API_KEY set?)"
             )
-            raise ValueError(f"Model {model!r} not found in {source}.{hint}")
+            raise ValueError(f"Model {requested_model!r} not found in {source}.{hint}")
 
     kwargs.setdefault("model", model_id)
     kwargs.setdefault("model_client", build_gateway_client(provider_name))

{hud_python-0.6.2 → hud_python-0.6.4}/hud/agents/openai_compatible/agent.py

@@ -193,16 +193,27 @@ class OpenAIChatAgent(ToolAgent[ChatCompletionMessageParam, OpenAIChatConfig]):
         sample: Sample | None = None
         if return_token_ids:
             prompt_token_ids = getattr(choice, "prompt_token_ids", None)
+            # Multimodal prompt (text + image chunks): the only prompt representation
+            # that survives image inputs; flat prompt_token_ids is null in that case.
+            prompt_chunks = getattr(choice, "prompt_chunks", None)
             token_ids = getattr(choice, "token_ids", None)
-            if prompt_token_ids is not None and token_ids is not None:
-                chat_state.continuation_token_ids = list(prompt_token_ids) + list(token_ids)
-                chat_state.continuation_message_count = len(messages)
+            has_prompt = prompt_token_ids is not None or prompt_chunks is not None
+            if token_ids is not None and has_prompt:
                 content_lp = choice.logprobs.content if choice.logprobs else None
                 sample = Sample(
-                    prompt_token_ids=list(prompt_token_ids),
+                    prompt_token_ids=list(prompt_token_ids) if prompt_token_ids is not None else [],
+                    prompt_chunks=list(prompt_chunks) if prompt_chunks is not None else None,
                     output_token_ids=list(token_ids),
                     output_logprobs=[tok.logprob for tok in content_lp] if content_lp else [],
                 )
+                # KV-cache continuation only applies to flat text prompts; clear any
+                # stale state when the gateway returns chunks-only (multimodal turn).
+                if prompt_token_ids is not None:
+                    chat_state.continuation_token_ids = list(prompt_token_ids) + list(token_ids)
+                    chat_state.continuation_message_count = len(messages)
+                else:
+                    chat_state.continuation_token_ids = None
+                    chat_state.continuation_message_count = None
 
         tool_calls: list[MCPToolCall] = []
         for tc in function_calls:

{hud_python-0.6.2 → hud_python-0.6.4}/hud/agents/tests/test_base.py

@@ -108,7 +108,7 @@ def test_create_agent_resolves_gateway_model_metadata(
 
     model = GatewayModelInfo(
         id="ft:custom-123",
-        model_name="gpt-5.4",
+        model_name="gpt-5.5",
         sdk_agent_type="openai_compatible",
         provider=GatewayProviderInfo(name="openai"),
     )
@@ -122,4 +122,40 @@ def test_create_agent_resolves_gateway_model_metadata(
     agent = create_agent("ft:custom-123")
 
     assert isinstance(agent, OpenAIChatAgent)
-    assert agent.config.model == "gpt-5.4"  # resolved to the model's real name
+    assert agent.config.model == "gpt-5.5"  # resolved to the model's real name
+
+
+@pytest.mark.parametrize(
+    ("alias", "canonical"),
+    [
+        ("deepseek-v4", "deepseek/deepseek-v4-pro"),
+        ("deepseek-v4-flash", "deepseek/deepseek-v4-flash"),
+        ("glm-5.2", "z-ai/glm-5.2"),
+        ("kimi-k2.6", "moonshotai/kimi-k2.6"),
+        ("minimax-m3", "MiniMax-M3"),
+    ],
+)
+def test_create_agent_accepts_gateway_model_aliases(
+    alias: str,
+    canonical: str,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    from hud.utils.gateway import GatewayModelInfo, GatewayProviderInfo
+
+    model = GatewayModelInfo(
+        id=canonical,
+        model_name=canonical,
+        sdk_agent_type="openai_compatible",
+        provider=GatewayProviderInfo(name="openai"),
+    )
+    monkeypatch.setattr("hud.agents.list_gateway_models", lambda: [model])
+
+    def _build_client(_provider: str) -> object:
+        return object()
+
+    monkeypatch.setattr("hud.agents.build_gateway_client", _build_client)
+
+    agent = create_agent(alias)
+
+    assert isinstance(agent, OpenAIChatAgent)
+    assert agent.config.model == canonical

{hud_python-0.6.2 → hud_python-0.6.4}/hud/agents/tests/test_provider_native_tools.py

@@ -102,7 +102,7 @@ def _commands(tool: Any) -> list[str]:
 
 
 async def test_openai_shell_wraps_command_with_timeout() -> None:
-    tool = OpenAIShellTool(spec=OpenAIShellTool.default_spec("gpt-5.4"), client=_ssh())
+    tool = OpenAIShellTool(spec=OpenAIShellTool.default_spec("gpt-5.5"), client=_ssh())
 
     result = await tool.execute({"commands": ["pwd"], "timeout_ms": 2500})
 
@@ -114,7 +114,7 @@ async def test_openai_shell_wraps_command_with_timeout() -> None:
 
 
 async def test_openai_shell_runs_each_command_without_timeout() -> None:
-    tool = OpenAIShellTool(spec=OpenAIShellTool.default_spec("gpt-5.4"), client=_ssh())
+    tool = OpenAIShellTool(spec=OpenAIShellTool.default_spec("gpt-5.5"), client=_ssh())
 
     await tool.execute({"commands": ["echo a", "echo b"]})
 
@@ -122,7 +122,7 @@ async def test_openai_shell_runs_each_command_without_timeout() -> None:
 
 
 async def test_openai_shell_rejects_non_list_commands_without_running() -> None:
-    tool = OpenAIShellTool(spec=OpenAIShellTool.default_spec("gpt-5.4"), client=_ssh())
+    tool = OpenAIShellTool(spec=OpenAIShellTool.default_spec("gpt-5.5"), client=_ssh())
 
     result = await tool.execute({"commands": 123})
 
@@ -131,7 +131,7 @@ async def test_openai_shell_rejects_non_list_commands_without_running() -> None:
 
 
 def test_openai_shell_to_params_is_shell_type() -> None:
-    tool = OpenAIShellTool(spec=OpenAIShellTool.default_spec("gpt-5.4"), client=_ssh())
+    tool = OpenAIShellTool(spec=OpenAIShellTool.default_spec("gpt-5.5"), client=_ssh())
     assert tool.to_params()["type"] == "shell"
 
 

{hud_python-0.6.2 → hud_python-0.6.4}/hud/agents/types.py

@@ -99,7 +99,7 @@ class OpenAIConfig(AgentConfig):
     """Configuration for OpenAIAgent."""
 
     model_name: str = "OpenAI"
-    model: str = Field(default="gpt-5.4", validation_alias=_model_alias)
+    model: str = Field(default="gpt-5.5", validation_alias=_model_alias)
     max_output_tokens: int | None = None
     temperature: float | None = None
     reasoning: Any = None  # openai Reasoning
@@ -113,7 +113,7 @@ class OpenAIChatConfig(AgentConfig):
     """Configuration for OpenAIChatAgent."""
 
     model_name: str = "OpenAI Chat"
-    model: str = Field(default="gpt-5-mini", validation_alias=_model_alias)
+    model: str = Field(default="gpt-5.4-mini", validation_alias=_model_alias)
     checkpoint: str | None = Field(
         default=None,
         description="Specific checkpoint name for inference routing. "
@@ -139,7 +139,7 @@ class ClaudeSDKConfig(AgentConfig):
     """
 
     model_name: str = "Claude Code"
-    model: str = Field(default="claude-sonnet-4-5", validation_alias=_model_alias)
+    model: str = Field(default="claude-sonnet-4-6", validation_alias=_model_alias)
     permission_mode: str = "bypassPermissions"
     max_steps: int = -1
     allowed_tools: list[str] = Field(
@@ -222,6 +222,10 @@ class Sample(BaseModel):
     """
 
     prompt_token_ids: list[int] = Field(default_factory=list[int])
+    # Multimodal prompt as serialized ``ModelInput`` chunks (text + image), set by
+    # vision rollouts where the prompt is not a flat token list. When present it is
+    # the authoritative prompt for training; ``prompt_token_ids`` stays empty.
+    prompt_chunks: list[dict[str, Any]] | None = None
     output_token_ids: list[int] = Field(default_factory=list[int])
     output_logprobs: list[float] = Field(default_factory=list[float])
 

{hud_python-0.6.2 → hud_python-0.6.4}/hud/cli/__init__.py

@@ -35,11 +35,13 @@ from .client import client_app  # noqa: E402
 from .deploy import deploy_command  # noqa: E402
 from .eval import eval_command  # noqa: E402
 from .init import init_command  # noqa: E402
+from .jobs import jobs_app  # noqa: E402
 from .login import login_command  # noqa: E402
 from .models import models_app  # noqa: E402
 from .serve import serve_command  # noqa: E402
 from .sync import sync_app  # noqa: E402
 from .task import task_app  # noqa: E402
+from .trace import trace_app  # noqa: E402
 
 app.command(name="serve")(serve_command)
 app.command(name="dev", deprecated=True, hidden=True)(serve_command)  # alias for now
@@ -49,6 +51,8 @@ app.command(name="eval")(eval_command)
 app.command(name="init")(init_command)
 app.command(name="cancel")(cancel_command)
 app.add_typer(models_app, name="models")
+app.add_typer(jobs_app, name="jobs")
+app.add_typer(trace_app, name="trace")
 
 
 @app.command(name="set")

{hud_python-0.6.2 → hud_python-0.6.4}/hud/cli/eval.py

@@ -5,6 +5,7 @@ Config Override Order: CLI arguments > .hud_eval.toml > defaults
 
 from __future__ import annotations
 
+import ast
 import asyncio
 import logging
 import os
@@ -42,8 +43,9 @@ def _resolve_model_from_catalog(model_id: str) -> tuple[AgentType, str] | None:
     Returns None if the model isn't found or the catalog is unreachable.
     """
     try:
-        from hud.utils.gateway import list_gateway_models
+        from hud.utils.gateway import list_gateway_models, normalize_gateway_model_id
 
+        model_id = normalize_gateway_model_id(model_id)
         models = list_gateway_models()
     except Exception:
         return None
@@ -116,8 +118,9 @@ class AgentPreset:
 
 _AGENT_PRESETS: list[AgentPreset] = [
     AgentPreset("Claude Sonnet 4.6", AgentType.CLAUDE, "claude-sonnet-4-6"),
-    AgentPreset("GPT-5.4", AgentType.OPENAI, "gpt-5.4"),
-    AgentPreset("Gemini 3.1 Pro (Preview)", AgentType.GEMINI, "gemini-3-1-pro"),
+    AgentPreset("Claude Opus 4.8", AgentType.CLAUDE, "claude-opus-4-8"),
+    AgentPreset("GPT-5.5", AgentType.OPENAI, "gpt-5.5"),
+    AgentPreset("Gemini 3.1 Pro (Preview)", AgentType.GEMINI, "gemini-3.1-pro-preview"),
     AgentPreset(
         "Grok 4-1 Fast (xAI)",
         AgentType.OPENAI_COMPATIBLE,
@@ -130,10 +133,22 @@ _AGENT_PRESETS: list[AgentPreset] = [
         },
     ),
     AgentPreset(
-        "GLM-4.6V (Z-AI)",
+        "GLM 5.2 (Z.ai)",
         AgentType.OPENAI_COMPATIBLE,
-        "z-ai/glm-4.6v",
-        {"openai_compatible": {"base_url": settings.hud_gateway_url, "model_name": "GLM-4.6V"}},
+        "z-ai/glm-5.2",
+        {"openai_compatible": {"base_url": settings.hud_gateway_url, "model_name": "GLM 5.2"}},
+    ),
+    AgentPreset(
+        "Kimi K2.6 (Moonshot)",
+        AgentType.OPENAI_COMPATIBLE,
+        "moonshotai/kimi-k2.6",
+        {"openai_compatible": {"base_url": settings.hud_gateway_url, "model_name": "Kimi K2.6"}},
+    ),
+    AgentPreset(
+        "MiniMax M3",
+        AgentType.OPENAI_COMPATIBLE,
+        "MiniMax-M3",
+        {"openai_compatible": {"base_url": settings.hud_gateway_url, "model_name": "MiniMax M3"}},
     ),
 ]
 
@@ -161,7 +176,7 @@ _DEFAULT_CONFIG_TEMPLATE = """# HUD Eval Configuration
 # use_computer_beta = true
 
 [openai]
-# model = "gpt-4o"
+# model = "gpt-5.5"
 # temperature = 0.7
 # max_output_tokens = 4096
 
@@ -401,6 +416,11 @@ class EvalConfig(BaseModel):
         if self.model:
             kwargs["model"] = self.model
 
+        if isinstance(kwargs.get("model"), str):
+            from hud.utils.gateway import normalize_gateway_model_id
+
+            kwargs["model"] = normalize_gateway_model_id(kwargs["model"])
+
         if self.agent_type == AgentType.OPENAI_COMPATIBLE and "api_key" not in kwargs:
             base_url = kwargs.get("base_url", "")
             if settings.hud_gateway_url in base_url and settings.api_key:
@@ -665,13 +685,46 @@ def _build_agent(cfg: EvalConfig) -> Any:
     return cast("Any", cfg.agent_type.cls)(config=config)
 
 
+def _python_defines_environment(path: Path) -> bool:
+    """Return True when ``path`` constructs a v6 :class:`~hud.environment.Environment`."""
+    try:
+        tree = ast.parse(path.read_text(encoding="utf-8"))
+    except (OSError, SyntaxError):
+        return False
+    for node in ast.walk(tree):
+        if not isinstance(node, ast.Call):
+            continue
+        callee = node.func
+        callee_name = (
+            callee.id
+            if isinstance(callee, ast.Name)
+            else callee.attr
+            if isinstance(callee, ast.Attribute)
+            else None
+        )
+        if callee_name == "Environment":
+            return True
+    return False
+
+
 def _spawn_target(source: Path) -> Path:
-    """The path the ``LocalRuntime`` provider serves: the source itself for ``.py``
-    files and directories, the surrounding directory for JSON/JSONL data files
-    (the env's ``.py`` source lives next to the tasks file)."""
+    """The path the ``LocalRuntime`` provider serves.
+
+    Directories and env-defining ``.py`` files are served as-is. Task-only
+    sources (``tasks.py`` importing from ``env.py``) resolve to a sibling
+    ``env.py`` or the containing directory. JSON/JSONL data files use the
+    surrounding directory (the env source lives next to the tasks file).
+    """
     resolved = source.resolve()
-    if resolved.is_dir() or resolved.suffix == ".py":
+    if resolved.is_dir():
+        return resolved
+    if resolved.suffix != ".py":
+        return resolved.parent
+    if _python_defines_environment(resolved):
         return resolved
+    env_py = resolved.parent / "env.py"
+    if env_py.is_file():
+        return env_py
     return resolved.parent
 
 

{hud_python-0.6.2 → hud_python-0.6.4}/hud/cli/init.py

@@ -76,8 +76,8 @@ def init_command(
         None,
         "--preset",
         "-p",
-        help="Starter preset to download from GitHub (e.g. blank, coding, browser, "
-        "deepresearch, rubrics, remote-browser). Omit for an interactive picker; in a "
+        help="Starter preset to download from GitHub (e.g. blank, browser, "
+        "deepresearch, cua, autonomous-businesses, verilog). Omit for an interactive picker; in a "
         "non-interactive shell, omitting it writes the minimal local scaffold.",
     ),
 ) -> None:
@@ -89,7 +89,7 @@ def init_command(
 
     Examples:
         hud init my-env                  # interactive picker (or local scaffold)
-        hud init my-env --preset coding  # download the coding starter
+        hud init my-env --preset browser  # download the browser starter
         hud init my-env --dir envs       # create ./envs/my-env[/not dim]
     """
     hud_console = HUDConsole()